drivers/thermal/power_allocator.c

   1 /*
   2  * A power allocator to manage temperature
   3  *
   4  * Copyright (C) 2014 ARM Ltd.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published by the Free Software Foundation.
   9  *
  10  * This program is distributed "as is" WITHOUT ANY WARRANTY of any
  11  * kind, whether express or implied; without even the implied warranty
  12  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
  13  * GNU General Public License for more details.
  14  */
  15
  16 #define pr_fmt(fmt) "Power allocator: " fmt
  17
  18 #include <linux/rculist.h>
  19 #include <linux/slab.h>
  20 #include <linux/thermal.h>
  21
  22 #define CREATE_TRACE_POINTS
  23 #include <trace/events/thermal_power_allocator.h>
  24
  25 #include "thermal_core.h"
  26
  27 #define FRAC_BITS 10
  28 #define int_to_frac(x) ((x) << FRAC_BITS)
  29 #define frac_to_int(x) ((x) >> FRAC_BITS)
  30
  31 /**
  32  * mul_frac() - multiply two fixed-point numbers
  33  * @x:  first multiplicand
  34  * @y:  second multiplicand
  35  *
  36  * Return: the result of multiplying two fixed-point numbers.  The
  37  * result is also a fixed-point number.
  38  */
  39 static inline s64 mul_frac(s64 x, s64 y)
  40 {
  41         return (x * y) >> FRAC_BITS;
  42 }
  43
  44 /**
  45  * div_frac() - divide two fixed-point numbers
  46  * @x:  the dividend
  47  * @y:  the divisor
  48  *
  49  * Return: the result of dividing two fixed-point numbers.  The
  50  * result is also a fixed-point number.
  51  */
  52 static inline s64 div_frac(s64 x, s64 y)
  53 {
  54         return div_s64(x << FRAC_BITS, y);
  55 }
  56
  57 /**
  58  * struct power_allocator_params - parameters for the power allocator governor
  59  * @err_integral:       accumulated error in the PID controller.
  60  * @prev_err:   error in the previous iteration of the PID controller.
  61  *              Used to calculate the derivative term.
  62  * @trip_switch_on:     first passive trip point of the thermal zone.  The
  63  *                      governor switches on when this trip point is crossed.
  64  * @trip_max_desired_temperature:       last passive trip point of the thermal
  65  *                                      zone.  The temperature we are
  66  *                                      controlling for.
  67  */
  68 struct power_allocator_params {
  69         s64 err_integral;
  70         s32 prev_err;
  71         int trip_switch_on;
  72         int trip_max_desired_temperature;
  73 };
  74
  75 /**
  76  * pid_controller() - PID controller
  77  * @tz: thermal zone we are operating in
  78  * @current_temp:       the current temperature in millicelsius
  79  * @control_temp:       the target temperature in millicelsius
  80  * @max_allocatable_power:      maximum allocatable power for this thermal zone
  81  *
  82  * This PID controller increases the available power budget so that the
  83  * temperature of the thermal zone gets as close as possible to
  84  * @control_temp and limits the power if it exceeds it.  k_po is the
  85  * proportional term when we are overshooting, k_pu is the
  86  * proportional term when we are undershooting.  integral_cutoff is a
  87  * threshold below which we stop accumulating the error.  The
  88  * accumulated error is only valid if the requested power will make
  89  * the system warmer.  If the system is mostly idle, there's no point
  90  * in accumulating positive error.
  91  *
  92  * Return: The power budget for the next period.
  93  */
  94 static u32 pid_controller(struct thermal_zone_device *tz,
  95                           unsigned long current_temp,
  96                           unsigned long control_temp,
  97                           u32 max_allocatable_power)
  98 {
  99         s64 p, i, d, power_range;
 100         s32 err, max_power_frac;
 101         struct power_allocator_params *params = tz->governor_data;
 102
 103         max_power_frac = int_to_frac(max_allocatable_power);
 104
 105         err = ((s32)control_temp - (s32)current_temp);
 106         err = int_to_frac(err);
 107
 108         /* Calculate the proportional term */
 109         p = mul_frac(err < 0 ? tz->tzp->k_po : tz->tzp->k_pu, err);
 110
 111         /*
 112          * Calculate the integral term
 113          *
 114          * if the error is less than cut off allow integration (but
 115          * the integral is limited to max power)
 116          */
 117         i = mul_frac(tz->tzp->k_i, params->err_integral);
 118
 119         if (err < int_to_frac(tz->tzp->integral_cutoff)) {
 120                 s64 i_next = i + mul_frac(tz->tzp->k_i, err);
 121
 122                 if (abs64(i_next) < max_power_frac) {
 123                         i = i_next;
 124                         params->err_integral += err;
 125                 }
 126         }
 127
 128         /*
 129          * Calculate the derivative term
 130          *
 131          * We do err - prev_err, so with a positive k_d, a decreasing
 132          * error (i.e. driving closer to the line) results in less
 133          * power being applied, slowing down the controller)
 134          */
 135         d = mul_frac(tz->tzp->k_d, err - params->prev_err);
 136         d = div_frac(d, tz->passive_delay);
 137         params->prev_err = err;
 138
 139         power_range = p + i + d;
 140
 141         /* feed-forward the known sustainable dissipatable power */
 142         power_range = tz->tzp->sustainable_power + frac_to_int(power_range);
 143
 144         power_range = clamp(power_range, (s64)0, (s64)max_allocatable_power);
 145
 146         trace_thermal_power_allocator_pid(tz, frac_to_int(err),
 147                                           frac_to_int(params->err_integral),
 148                                           frac_to_int(p), frac_to_int(i),
 149                                           frac_to_int(d), power_range);
 150
 151         return power_range;
 152 }
 153
 154 /**
 155  * divvy_up_power() - divvy the allocated power between the actors
 156  * @req_power:  each actor's requested power
 157  * @max_power:  each actor's maximum available power
 158  * @num_actors: size of the @req_power, @max_power and @granted_power's array
 159  * @total_req_power: sum of @req_power
 160  * @power_range:        total allocated power
 161  * @granted_power:      output array: each actor's granted power
 162  * @extra_actor_power:  an appropriately sized array to be used in the
 163  *                      function as temporary storage of the extra power given
 164  *                      to the actors
 165  *
 166  * This function divides the total allocated power (@power_range)
 167  * fairly between the actors.  It first tries to give each actor a
 168  * share of the @power_range according to how much power it requested
 169  * compared to the rest of the actors.  For example, if only one actor
 170  * requests power, then it receives all the @power_range.  If
 171  * three actors each requests 1mW, each receives a third of the
 172  * @power_range.
 173  *
 174  * If any actor received more than their maximum power, then that
 175  * surplus is re-divvied among the actors based on how far they are
 176  * from their respective maximums.
 177  *
 178  * Granted power for each actor is written to @granted_power, which
 179  * should've been allocated by the calling function.
 180  */
 181 static void divvy_up_power(u32 *req_power, u32 *max_power, int num_actors,
 182                            u32 total_req_power, u32 power_range,
 183                            u32 *granted_power, u32 *extra_actor_power)
 184 {
 185         u32 extra_power, capped_extra_power;
 186         int i;
 187
 188         /*
 189          * Prevent division by 0 if none of the actors request power.
 190          */
 191         if (!total_req_power)
 192                 total_req_power = 1;
 193
 194         capped_extra_power = 0;
 195         extra_power = 0;
 196         for (i = 0; i < num_actors; i++) {
 197                 u64 req_range = req_power[i] * power_range;
 198
 199                 granted_power[i] = DIV_ROUND_CLOSEST_ULL(req_range,
 200                                                          total_req_power);
 201
 202                 if (granted_power[i] > max_power[i]) {
 203                         extra_power += granted_power[i] - max_power[i];
 204                         granted_power[i] = max_power[i];
 205                 }
 206
 207                 extra_actor_power[i] = max_power[i] - granted_power[i];
 208                 capped_extra_power += extra_actor_power[i];
 209         }
 210
 211         if (!extra_power)
 212                 return;
 213
 214         /*
 215          * Re-divvy the reclaimed extra among actors based on
 216          * how far they are from the max
 217          */
 218         extra_power = min(extra_power, capped_extra_power);
 219         if (capped_extra_power > 0)
 220                 for (i = 0; i < num_actors; i++)
 221                         granted_power[i] += (extra_actor_power[i] *
 222                                         extra_power) / capped_extra_power;
 223 }
 224
 225 static int allocate_power(struct thermal_zone_device *tz,
 226                           unsigned long current_temp,
 227                           unsigned long control_temp)
 228 {
 229         struct thermal_instance *instance;
 230         struct power_allocator_params *params = tz->governor_data;
 231         u32 *req_power, *max_power, *granted_power, *extra_actor_power;
 232         u32 *weighted_req_power;
 233         u32 total_req_power, max_allocatable_power, total_weighted_req_power;
 234         u32 total_granted_power, power_range;
 235         int i, num_actors, total_weight, ret = 0;
 236         int trip_max_desired_temperature = params->trip_max_desired_temperature;
 237
 238         mutex_lock(&tz->lock);
 239
 240         num_actors = 0;
 241         total_weight = 0;
 242         list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
 243                 if ((instance->trip == trip_max_desired_temperature) &&
 244                     cdev_is_power_actor(instance->cdev)) {
 245                         num_actors++;
 246                         total_weight += instance->weight;
 247                 }
 248         }
 249
 250         /*
 251          * We need to allocate five arrays of the same size:
 252          * req_power, max_power, granted_power, extra_actor_power and
 253          * weighted_req_power.  They are going to be needed until this
 254          * function returns.  Allocate them all in one go to simplify
 255          * the allocation and deallocation logic.
 256          */
 257         BUILD_BUG_ON(sizeof(*req_power) != sizeof(*max_power));
 258         BUILD_BUG_ON(sizeof(*req_power) != sizeof(*granted_power));
 259         BUILD_BUG_ON(sizeof(*req_power) != sizeof(*extra_actor_power));
 260         BUILD_BUG_ON(sizeof(*req_power) != sizeof(*weighted_req_power));
 261         req_power = devm_kcalloc(&tz->device, num_actors * 5,
 262                                  sizeof(*req_power), GFP_KERNEL);
 263         if (!req_power) {
 264                 ret = -ENOMEM;
 265                 goto unlock;
 266         }
 267
 268         max_power = &req_power[num_actors];
 269         granted_power = &req_power[2 * num_actors];
 270         extra_actor_power = &req_power[3 * num_actors];
 271         weighted_req_power = &req_power[4 * num_actors];
 272
 273         i = 0;
 274         total_weighted_req_power = 0;
 275         total_req_power = 0;
 276         max_allocatable_power = 0;
 277
 278         list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
 279                 int weight;
 280                 struct thermal_cooling_device *cdev = instance->cdev;
 281
 282                 if (instance->trip != trip_max_desired_temperature)
 283                         continue;
 284
 285                 if (!cdev_is_power_actor(cdev))
 286                         continue;
 287
 288                 if (cdev->ops->get_requested_power(cdev, tz, &req_power[i]))
 289                         continue;
 290
 291                 if (!total_weight)
 292                         weight = 1 << FRAC_BITS;
 293                 else
 294                         weight = instance->weight;
 295
 296                 weighted_req_power[i] = frac_to_int(weight * req_power[i]);
 297
 298                 if (power_actor_get_max_power(cdev, tz, &max_power[i]))
 299                         continue;
 300
 301                 total_req_power += req_power[i];
 302                 max_allocatable_power += max_power[i];
 303                 total_weighted_req_power += weighted_req_power[i];
 304
 305                 i++;
 306         }
 307
 308         power_range = pid_controller(tz, current_temp, control_temp,
 309                                      max_allocatable_power);
 310
 311         divvy_up_power(weighted_req_power, max_power, num_actors,
 312                        total_weighted_req_power, power_range, granted_power,
 313                        extra_actor_power);
 314
 315         total_granted_power = 0;
 316         i = 0;
 317         list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
 318                 if (instance->trip != trip_max_desired_temperature)
 319                         continue;
 320
 321                 if (!cdev_is_power_actor(instance->cdev))
 322                         continue;
 323
 324                 power_actor_set_power(instance->cdev, instance,
 325                                       granted_power[i]);
 326                 total_granted_power += granted_power[i];
 327
 328                 i++;
 329         }
 330
 331         trace_thermal_power_allocator(tz, req_power, total_req_power,
 332                                       granted_power, total_granted_power,
 333                                       num_actors, power_range,
 334                                       max_allocatable_power, current_temp,
 335                                       (s32)control_temp - (s32)current_temp);
 336
 337         kfree(req_power);
 338 unlock:
 339         mutex_unlock(&tz->lock);
 340
 341         return ret;
 342 }
 343
 344 static int get_governor_trips(struct thermal_zone_device *tz,
 345                               struct power_allocator_params *params)
 346 {
 347         int i, ret, last_passive;
 348         bool found_first_passive;
 349
 350         found_first_passive = false;
 351         last_passive = -1;
 352         ret = -EINVAL;
 353
 354         for (i = 0; i < tz->trips; i++) {
 355                 enum thermal_trip_type type;
 356
 357                 ret = tz->ops->get_trip_type(tz, i, &type);
 358                 if (ret)
 359                         return ret;
 360
 361                 if (!found_first_passive) {
 362                         if (type == THERMAL_TRIP_PASSIVE) {
 363                                 params->trip_switch_on = i;
 364                                 found_first_passive = true;
 365                         }
 366                 } else if (type == THERMAL_TRIP_PASSIVE) {
 367                         last_passive = i;
 368                 } else {
 369                         break;
 370                 }
 371         }
 372
 373         if (last_passive != -1) {
 374                 params->trip_max_desired_temperature = last_passive;
 375                 ret = 0;
 376         } else {
 377                 ret = -EINVAL;
 378         }
 379
 380         return ret;
 381 }
 382
 383 static void reset_pid_controller(struct power_allocator_params *params)
 384 {
 385         params->err_integral = 0;
 386         params->prev_err = 0;
 387 }
 388
 389 static void allow_maximum_power(struct thermal_zone_device *tz)
 390 {
 391         struct thermal_instance *instance;
 392         struct power_allocator_params *params = tz->governor_data;
 393
 394         list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
 395                 if ((instance->trip != params->trip_max_desired_temperature) ||
 396                     (!cdev_is_power_actor(instance->cdev)))
 397                         continue;
 398
 399                 instance->target = 0;
 400                 instance->cdev->updated = false;
 401                 thermal_cdev_update(instance->cdev);
 402         }
 403 }
 404
 405 /**
 406  * power_allocator_bind() - bind the power_allocator governor to a thermal zone
 407  * @tz: thermal zone to bind it to
 408  *
 409  * Check that the thermal zone is valid for this governor, that is, it
 410  * has two thermal trips.  If so, initialize the PID controller
 411  * parameters and bind it to the thermal zone.
 412  *
 413  * Return: 0 on success, -EINVAL if the trips were invalid or -ENOMEM
 414  * if we ran out of memory.
 415  */
 416 static int power_allocator_bind(struct thermal_zone_device *tz)
 417 {
 418         int ret;
 419         struct power_allocator_params *params;
 420         unsigned long switch_on_temp, control_temp;
 421         u32 temperature_threshold;
 422
 423         if (!tz->tzp || !tz->tzp->sustainable_power) {
 424                 dev_err(&tz->device,
 425                         "power_allocator: missing sustainable_power\n");
 426                 return -EINVAL;
 427         }
 428
 429         params = kzalloc(sizeof(*params), GFP_KERNEL);
 430         if (!params)
 431                 return -ENOMEM;
 432
 433         ret = get_governor_trips(tz, params);
 434         if (ret) {
 435                 dev_err(&tz->device,
 436                         "thermal zone %s has wrong trip setup for power allocator\n",
 437                         tz->type);
 438                 goto free;
 439         }
 440
 441         ret = tz->ops->get_trip_temp(tz, params->trip_switch_on,
 442                                      &switch_on_temp);
 443         if (ret)
 444                 goto free;
 445
 446         ret = tz->ops->get_trip_temp(tz, params->trip_max_desired_temperature,
 447                                      &control_temp);
 448         if (ret)
 449                 goto free;
 450
 451         temperature_threshold = control_temp - switch_on_temp;
 452
 453         tz->tzp->k_po = tz->tzp->k_po ?:
 454                 int_to_frac(tz->tzp->sustainable_power) / temperature_threshold;
 455         tz->tzp->k_pu = tz->tzp->k_pu ?:
 456                 int_to_frac(2 * tz->tzp->sustainable_power) /
 457                 temperature_threshold;
 458         tz->tzp->k_i = tz->tzp->k_i ?: int_to_frac(10) / 1000;
 459         /*
 460          * The default for k_d and integral_cutoff is 0, so we can
 461          * leave them as they are.
 462          */
 463
 464         reset_pid_controller(params);
 465
 466         tz->governor_data = params;
 467
 468         return 0;
 469
 470 free:
 471         kfree(params);
 472         return ret;
 473 }
 474
 475 static void power_allocator_unbind(struct thermal_zone_device *tz)
 476 {
 477         dev_dbg(&tz->device, "Unbinding from thermal zone %d\n", tz->id);
 478         kfree(tz->governor_data);
 479         tz->governor_data = NULL;
 480 }
 481
 482 static int power_allocator_throttle(struct thermal_zone_device *tz, int trip)
 483 {
 484         int ret;
 485         unsigned long switch_on_temp, control_temp, current_temp;
 486         struct power_allocator_params *params = tz->governor_data;
 487
 488         /*
 489          * We get called for every trip point but we only need to do
 490          * our calculations once
 491          */
 492         if (trip != params->trip_max_desired_temperature)
 493                 return 0;
 494
 495         ret = thermal_zone_get_temp(tz, &current_temp);
 496         if (ret) {
 497                 dev_warn(&tz->device, "Failed to get temperature: %d\n", ret);
 498                 return ret;
 499         }
 500
 501         ret = tz->ops->get_trip_temp(tz, params->trip_switch_on,
 502                                      &switch_on_temp);
 503         if (ret) {
 504                 dev_warn(&tz->device,
 505                          "Failed to get switch on temperature: %d\n", ret);
 506                 return ret;
 507         }
 508
 509         if (current_temp < switch_on_temp) {
 510                 tz->passive = 0;
 511                 reset_pid_controller(params);
 512                 allow_maximum_power(tz);
 513                 return 0;
 514         }
 515
 516         tz->passive = 1;
 517
 518         ret = tz->ops->get_trip_temp(tz, params->trip_max_desired_temperature,
 519                                 &control_temp);
 520         if (ret) {
 521                 dev_warn(&tz->device,
 522                          "Failed to get the maximum desired temperature: %d\n",
 523                          ret);
 524                 return ret;
 525         }
 526
 527         return allocate_power(tz, current_temp, control_temp);
 528 }
 529
 530 static struct thermal_governor thermal_gov_power_allocator = {
 531         .name           = "power_allocator",
 532         .bind_to_tz     = power_allocator_bind,
 533         .unbind_from_tz = power_allocator_unbind,
 534         .throttle       = power_allocator_throttle,
 535 };
 536
 537 int thermal_gov_power_allocator_register(void)
 538 {
 539         return thermal_register_governor(&thermal_gov_power_allocator);
 540 }
 541
 542 void thermal_gov_power_allocator_unregister(void)
 543 {
 544         thermal_unregister_governor(&thermal_gov_power_allocator);
 545 }